Nvidia reportedly plans to invest in d-Matrix, whose Raptor inference chips are designed to integrate with Nvidia MGX racks via NVLink Fusion; the post says racks are expected in Q4 2027. d-Matrix reports 0.37 pJ/bit on early silicon and targets 100+ TB/s per card, compared with an HBM4 reference of about 18 TB/s at 2–3 pJ/bit.